Evaluation of performance disparities in deep learning models for neuroimaging and the role of entropy-based active learning

Abstract

Deep learning models are now the leading approach for medical image segmentation. However, they can produce biased predictions across demographic groups such as sex and race. These disparities are important because they can affect downstream analyses, including volumetric studies, and may contribute to unequal clinical outcomes. This thesis addresses this issue by studying demographic bias in structural brain MRI segmentation and by proposing a fairness-aware active learning framework to reduce these disparities. In the first study, published in the Machine Learning for Biomedical Imaging journal, we compared several deep learning models, including nnU-Net, UNesT, and CoTr, as well as a traditional atlas-based method, ANTs. Using manually curated gold-standard segmentations of the nucleus accumbens, we showed that some models are much more sensitive than others to demographic imbalance in the training data. We also found that biased automated segmentations can obscure true race-related volumetric differences that remain visible in the manual annotations, thereby influencing downstream morphometric conclusions. In the second study, which will be presented at the 2026 Medical Imaging with Deep Learning conference, we introduced a fairness-aware active learning strategy to reduce these disparities during sample selection. The proposed weighted localized entropy method combines group-specific performance weighting with localized uncertainty estimation. By focusing uncertainty within a region of interest, the method better captures meaningful epistemic uncertainty and avoids simple anatomical variation. Experiments on synthetic MRI data with controlled morphological bias showed that this strategy can substantially reduce group disparity, with improvements of up to 86% relative to standard entropy sampling. Overall, this work shows that fairness in neuroimaging should not be treated as a secondary evaluation criterion. It should be considered a central design principle. By identifying bias in brain segmentation across both deep learning models and traditional methods, and by proposing a practical strategy for training more equitable models, this thesis advances the development of more reliable and fair medical image analysis.

Date	30 Apr 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Sylvain Bouix (Supervisor)